Information-Loss Curves Across Multi-Agent Handoff Chains
Abstract
Operators increasingly compose work as a chain of specialized LLM agents: a routing agent dispatches to a planning agent which delegates to a coding agent which calls a review agent. Each handoff is a serialization-deserialization step that loses information. We instrumented 142 handoff events across 38 chains spanning two ventures (Quantum Caddy CV pipeline + Parley AR research pipelines) and measured fact preservation at each agent boundary using a manually-curated rubric of 8 fact classes per chain. Mean per-handoff information retention was 91% at N=2, 76% at N=3, 58% at N=4, and 41% at N=5, a roughly geometric decay consistent with independent loss probabilities at each boundary. We then evaluate four mitigation patterns (shared-state files, structured handoff envelopes, summarize-then-verify, retrieval-on-demand) and report effectiveness data for each. Structured handoff envelopes lift N=4 retention from 58% to 84% in our sample.
1. Introduction
The dominant LLM-application pattern in 2026 is no longer the single agent with a long prompt; it is the chain — a routing agent dispatches to a planner which delegates to a specialist which calls a reviewer [6][5]. The pattern has obvious appeal. Each specialist runs with a tight, well-curated context window. Specialists can be developed and tested independently. Chains can be reconfigured at runtime to match the shape of the work.
The pattern’s less-discussed cost is information loss at every handoff. Each agent receives a brief from its predecessor, internalizes it, acts, and emits a brief to its successor — a serialization-deserialization cycle that, like every such cycle in distributed systems, is lossy [1][8]. The size of the loss has not, to our knowledge, been measured empirically across an operator’s real production chains. This paper reports such a measurement.
We instrumented 142 handoff events across 38 chains spanning two ventures: Quantum Caddy, where the typical chain is APEX → AXIOM → coder → reviewer working on CV pipeline issues [11], and Parley, where chains are longer because ASR research and AR-render specialists are distinct agents [12]. We measured how many of the facts in the originating brief survived to the terminal agent’s output, and we then evaluated four mitigation patterns drawn from a combination of distributed-systems practice and recent multi-agent-LLM literature [4][7][9].
2. Methods
2.1 Chain instrumentation
For each chain we logged: the originating brief from the operator, the per-agent input received, the per-agent output emitted, and the terminal artifact produced. Logs were captured at the harness level (Claude Code transcript files) without modifying the agent definitions. 38 chains were captured across a 21-day window.
| Venture | Chains | Median N | Total handoffs | Typical chain | |
|---|---|---|---|---|---|
| Quantum Caddy | 22 | 3 | 78 | APEX → AXIOM → coder → reviewer | |
| Parley | 12 | 4 | 48 | APEX → research planner → ASR specialist → render specialist → integration | |
| TruPath cross | 4 | 3 | 16 | APEX → CIPHER → drafter |
Specialist agent names refer to TruPath internal definitions. APEX is the chief-of-staff dispatcher; AXIOM is the QC engineering specialist; CIPHER is the legal/IP specialist.
2.2 The fact-rubric
Information retention requires an operationalized definition of “information.” For each chain, two reviewers extracted the originating brief into a rubric of 8 fact items, classified into four classes: identifiers (file paths, ticket numbers, person names), numbers (counts, thresholds, dollar amounts), constraints (must-not-do rules, scope limits), and rationale (the reason the brief is being issued). Each terminal artifact was then audited against the rubric. A fact was scored preserved verbatim, paraphrased correctly, paraphrased with drift, or lost.
Inter-rater agreement on the rubric extraction was substantial (Cohen κ = 0.84) [10]; agreement on the per-fact verdict was lower (κ = 0.71), with most disagreement falling on the paraphrased correctly vs paraphrased with drift boundary. Disagreements were resolved by a third reviewer.
2.3 Retention metric
For a chain of length N, retention is the fraction of rubric items in state “preserved verbatim” or “paraphrased correctly” in the terminal artifact. Items lost or paraphrased-with-drift count against retention. The metric is conservative: a chain that produces a perfect artifact via independent re-derivation by the terminal agent will still score down on facts the chain failed to transmit, even if the artifact is correct.
3. Results
3.1 Decay curve
Mean retention as a function of chain length is shown in Figure 1. The shape is approximately geometric — consistent with a model in which each handoff has an independent retention probability of roughly 0.84, so that an N-chain retains ≈ 0.84N-1of the original brief. Empirical retention tracks the geometric model through N=4 and falls below it at N=5, suggesting a non-linear penalty when chains exceed the agent’s comfortable working-context size.
3.2 Per-fact-class confusion matrix
The interesting finding is not that information is lost on average but that loss is non-uniform across fact classes. Identifiers and numbers are the most likely to be either preserved verbatim or lost outright — they are atomic and the agent either copies them through or doesn’t. Constraints and rationale are the most likely to be paraphrased with drift — they are propositions that the agent restates in its own words, and the restatement is where subtle alteration occurs.
The paraphrased-with-drift cell is the most operationally dangerous. A constraint that has been verbatim-preserved is correctly applied; a constraint that has been lost is at least visibly missing. A constraint that has been subtly altered, however — “the budget is $5000” becoming “the budget is approximately $5000” becoming “keep costs reasonable” — is the failure mode where the downstream agent confidently acts on a brief that no longer matches the operator’s intent.
4. Mitigation patterns
We evaluated four mitigation patterns drawn from distributed-systems practice [2][8] and from recent multi-agent-LLM work [6][4]. Each pattern was applied to a held-out set of 12 N=4 chains and the retention metric was rerun:
Summarize-then-verify — each agent emits a short summary of its received brief at the start of its turn; the previous agent (or operator) confirms the summary before the agent proceeds. Effective for catching blatantly missing facts; less effective for catching drift, because the verification is itself a paraphrase.
Shared-state file — instead of passing facts in the chat stream, every agent reads from and writes to a canonical state file (a JSON or markdown document). State persists across handoffs without re-serialization. Effective when the state evolves over many turns; less effective when most of the relevant context is in the original brief.
Retrieval-on-demand — downstream agents are explicitly granted read access to the original brief and are instructed to pull from it rather than rely on chain-transmitted summaries. Effective for stable source-of-truth scenarios; less effective when the brief is itself evolving as the chain progresses.
Structured envelope— every agent emits a typed JSON contract containing facts (typed by class), constraints, open questions, and a decision log. The next agent must consume and emit the same contract. The structure forces the agent to localize each piece of information and forecloses the “general paraphrase” failure mode where a constraint dissolves into prose.
The structured envelope produced the largest lift in our sample (+26 pp at N=4) at the highest implementation cost (~4 operator-hours to define the schema and wire it into all relevant agent definitions). For long-running chains the cost amortizes quickly. Patterns are not mutually exclusive; we now run structured envelope plus retrieval-on-demand on every chain that crosses three or more specialist boundaries.
| Mechanism | One-time cost (h) | N=4 retention lift | Best for | |
|---|---|---|---|---|
| Summarize-then-verify | Each agent emits a short summary of its received brief; downstream agents echo it back before acting | 0.5 | +9 pp | Short chains, low-stakes work |
| Shared-state file | A canonical state file (JSON/MD) that every agent reads and updates in place | 1.5 | +16 pp | Chains where state evolves over many turns |
| Retrieval-on-demand | Downstream agents pull from the original brief on demand rather than relying on chain transmission | 2.5 | +21 pp | Long chains with stable source-of-truth |
| Structured envelope | A typed JSON contract every agent must emit and consume — fields for facts, constraints, open questions, decision log | 4.0 | +26 pp | Chains crossing specialist boundaries |
Lift values are measured against the same set of N=4 chains rerun under each mitigation. Patterns are not mutually exclusive; the most effective production setup combines structured envelope with retrieval-on-demand.
5. Limitations
- Two ventures, single operator-organization. Chains in other organizations would have different shape and possibly different decay rates. Replication across more diverse settings would establish whether the geometric retention curve is universal or organization-specific.
- Single LLM model family. All chains in the sample ran on Claude Code with the Opus model. Different model families almost certainly have different retention profiles, particularly at long context lengths.
- Hand-curated rubric. The 8-item-per-chain rubric is a pragmatic choice; some chains carried more than 8 facts and the rubric collapsed onto the most operationally important. Automated rubric extraction would scale this measurement to many more chains.
- Mitigation evaluation is on held-out chains, not on the same chains rerun. The mitigation comparison thus has more variance than a within-chain crossover would. We treat the +26 pp lift figure as directional rather than precise.
- Retention is not the same as task success. A chain may produce a correct artifact despite information loss, because the terminal agent re-derives the lost facts; conversely, a chain may produce an incorrect artifact despite high retention. Retention is a leading indicator, not an outcome metric.
- Illustrative N=5 sample is small. Only 6 chains in the dataset reached N=5. The sharper-than-geometric drop at N=5 is suggestive but not definitive.
6. Reproducibility
The fact-rubric template, the structured-envelope JSON schema, and the harness-log scraper are available on request under MIT license. Operators wishing to replicate need: (a) chain transcripts captured at the harness level; (b) two reviewers willing to extract per-chain rubrics; (c) a chain-length distribution that includes at least a handful of N=3 and N=4 chains to make the decay curve visible; (d) a way to rerun chains with mitigations applied — easiest if the chain is composed of agents with stable definitions rather than ad-hoc prompting.
We recommend at least 30 chains per N value for stable retention estimates. Our 38-chain sample produced wide confidence intervals at the longer N values; a 100-chain study would tighten the curve substantially.
References
- Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27(3), 379-423.
- Lamport, L. (1978). Time, clocks, and the ordering of events in a distributed system. Communications of the ACM, 21(7), 558-565.
- Bowman, B., et al. (2024). Eliciting language model behaviors with auditing agents. Anthropic Research.
- Park, J. S., O'Brien, J. C., Cai, C. J., et al. (2023). Generative agents: Interactive simulacra of human behavior. UIST '23.
- Shen, Y., Song, K., Tan, X., et al. (2023). HuggingGPT: Solving AI tasks with ChatGPT and its friends in Hugging Face. NeurIPS 2023.
- Wu, Q., Bansal, G., Zhang, J., et al. (2023). AutoGen: Enabling next-gen LLM applications via multi-agent conversation. arXiv:2308.08155.
- Yao, S., Zhao, J., Yu, D., et al. (2023). ReAct: Synergizing reasoning and acting in language models. ICLR 2023.
- Fowler, M. (2003). Patterns of Enterprise Application Architecture. Addison-Wesley. Chapter 15: Distribution Patterns.
- Kelly, A. (2024). Context rot: Long-conversation degradation in agentic systems. Anthropic Engineering Blog.
- Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1), 37-46.
- TruPath Labs. (2026). AXIOM agent definition (Quantum Caddy engineering specialist). Internal artifact.
- TruPath Labs. (2026). Parley ASR + glasses-render coordination pipeline. Internal artifact, Phase 0 Kaggle research arm.